1354697 - Differences in Ventricular Volumetric Measurements and Prognostication Based on Basal Slice Interpolation by Deep Learning Algorithms

Deep learning algorithms allow for accurate and reproducible volumetric measurements in cardiac magnetic resonance (CMR) imaging. However, segmentation of the basal slices in the ventricle often leads to inconsistent measurements when using the summation of disks method. The goal of this study was to compare ventricular volumetric measurements with and without insertion of mitral and tricuspid annulus lines for basal slice interpolation.

Methods:

Two hundred patients were identified from a registry of individuals who underwent vasodilator stress CMR. The left ventricle (LV) and right ventricle (RV) were contoured with and without annulus lines using a fully automated deep learning algorithm by SuiteHeart (Neosoft). End-diastolic volume (EDV), end-systolic volume (ESV), ejection fraction (EF), LV mass, early LV peak filling rate (PFR), time to PFR, and time to fill 80% of the LV were measured. Patients were followed for a minimum of 3 years after CMR examination by reviewing medical records and/or by telephone contacts. Clinical outcome was a composite of all-cause death, hospital admission for ventricular arrhythmia or heart failure, LV assist device implantation, and heart transplantation. Ventricular volumetric measurements were compared using paired t-test. Agreement was assessed with Bland-Altman plots. Category-free net reclassification index was used to evaluate for improvement in risk prediction based on time-to-event data. Significance was determined based on p < 0.05.

Results:

Automated segmentation of the RV and LV with annulus lines at end-diastole and end-systole is demonstrated in Figure 1. Biventricular measurements are shown in Table 1. Inclusion of the annulus lines resulted in significantly higher LV ESV (100.0 ± 58.3 vs 95.8 ± 54.5 ml, p = 0.03), RV ESV (78.2 ± 34.1 vs 72.7 ± 34.7 ml, p < 0.001) and LV mass (135.0 ± 71.1 vs 132.7 ± 71.5 g, p < 0.001). There was a significantly lower LV EF (50.0 ± 13.0 vs 51.3 ± 13.8 %, p < 0.001), RV EF (55.3 ± 10.7 vs 57.9 ± 11.2 %, p < 0.001), early PFR (448.3 ± 160.5 vs 618.6 ± 284.0 ml/s, p < 0.001), and time to PFR (154.7 ± 57.3 vs 195.5 ± 146.0 ms, p < 0.001) with inclusion of annulus lines. The limits of agreement are shown in Figure 2. Among the significantly different parameters, there were no improvements in predicting clinical outcomes based on net reclassification identification with the addition of the annulus lines.  

Conclusion:

Automated segmentation with annulus lines for basal slice interpolation results in statistically significant different measurements of ventricular size, function, and relaxation. However, these differences do not impact their ability to predict clinical outcomes.